A TWO-STAGE DEEP LEARNING FRAMEWORK FOR AUTOMATED GLAUCOMA DETECTION USING U-NET AND CNN ENSEMBLES

Abstract

Glaucoma is a leading cause of irreversible blindness worldwide, often progressing asymptomatically until advanced stages, making early detection critical yet challenging. Conventional screening methods are largely subjective, time-intensive, and reliant on skilled ophthalmologists, whose availability is limited in densely populated regions such as India. To address these limitations, we propose a two-stage automated glaucoma screening system. In the first stage, the optic disc is segmented using a U-Net-based deep learning architecture. In the second stage, glaucoma classification is performed using both pretrained deep convolutional neural networks (CNNs) and a customized CNN model. Despite existing advancements, challenges such as inconsistent image quality, segmentation errors affecting cup-to-disc ratio (CDR) estimation, and limited model generalizability persist. This study addresses these gaps by improving segmentation accuracy and classification precision through an ensemble learning approach that integrates both pretrained and customized CNN models. Experimental evaluation on the DRISHTI-GS1 dataset demonstrates that the proposed ensemble model achieves a classification accuracy of 96%, outperforming conventional methods. Additionally, the CDR computed from the segmented optic cup and disc serves as a valuable reference metric to support early-stage glaucoma diagnosis by clinicians.